Erdős–Rényi Random Graphs

• Erdős–Rényi graphs are random graphs constructed with n vertices and edge probability p, forming a foundational model in probabilistic combinatorics.
• They exhibit sharp phase transitions such as the emergence of a giant component and connectivity thresholds, relevant to network science and statistical physics.
• Their analysis leverages degree distribution approximations, spectral theory, and large-deviation principles to guide algorithmic and inferential methods.

The Erdős–Rényi graph, denoted $G(n, p)$, is the canonical model of a random undirected simple graph on $n$ labelled vertices wherein each pair is connected by an edge independently with probability $p$. The model provides a foundational framework for probabilistic combinatorics and random graph theory, underpinning structural, spectral, algorithmic, and probabilistic analysis across a range of regimes and applications. Both finite Erdős–Rényi graphs and their infinite (countable) limit play critical roles in graph theory, statistical physics, combinatorics, and network science.

1. Model Definitions and Basic Properties

Two standard ensembles define the Erdős–Rényi graph:

• $G(n, p)$: Each of the $\binom{n}{2}$ possible edges is present independently with fixed probability $p \in (0,1)$.
• $G(n, m)$: A graph chosen uniformly at random from the set of all graphs with $n$ vertices and exactly $m$ edges.

For a fixed $p$, as $n \to \infty$, $G(n,p)$ almost surely exhibits sharp threshold phenomena, including emergence of connectivity, appearance of subgraphs, and phase transitions in components. The expected number of edges is $p\binom{n}{2}$, and the degree $D_i$ of vertex $i$ follows

$$D_i \sim \mathrm{Bin}(n-1, p), \quad \mathbb{E}[D_i] = (n-1)p, \quad \operatorname{Var}(D_i) = (n-1)p(1-p)$$

(Alarfaj et al., 2023).

For large $n$, the degree distribution can be approximated by a normal distribution if $np$ and $n(1-p)$ are both at least 5; a sharper rule for multivariate normality is $np \gtrsim 10$.

In the sparse regime ($p = c/n$ with constant $c > 0$), the expected degree converges to $c$, with limiting local structure described by a Galton–Watson tree with Poisson($c$) offspring distribution (Jung et al., 2017).

Infinite Limit and the Rado Graph

As $n \to \infty$ with constant $p \in (0,1)$, the limiting object is the countable random graph $R$, characterized uniquely (up to isomorphism) by the extension property: for any finite disjoint $U, V \subset \mathbb{N}$, there is a vertex joined to all vertices in $U$ and none in $V$. This object, also called the Rado graph, is homogeneous, universal, and self-complementary, and its automorphism group is simple, transitive on finite configurations, and of cardinality $2^{\aleph_0}$ (Cameron, 2013). All first-order properties obey a zero–one law: for every first-order sentence $\varphi$, $\lim_{n\to\infty} P[G(n,p) \models \varphi]=0$ or 1.

2. Degree Distributions, Correlations, and Inference

The structure of node degrees in $G(n,p)$ underpins many inferences:

• Covariance between distinct degrees is $\mathrm{Cov}(D_i, D_j) = p(1-p)$, and correlation $\frac{1}{n-1}$, reflecting weak dependence (Alarfaj et al., 2023).
• The joint degree distribution for large $n$ is well-approximated by an $n$-variate normal,

$$\mathbf{D} \approx \mathcal{N}_n\left(\mu, \Sigma\right)$$

with $\boldsymbol{\mu} = (n-1)p \, \mathbf{1}_n$, $\Sigma_{ii} = (n-1)p(1-p)$, $\Sigma_{ij}=p(1-p)$ for $i\ne j$. The approximation is valid for $np \gtrsim 10$.

Goodness-of-fit tests (chi-square, Anderson–Darling) confirm the multivariate normality for large $n, np$, but not for small parameters. Maximum likelihood estimation of $p$ under this model achieves minimal mean squared error for the dependent-degree MLE (Alarfaj et al., 2023).

3. Phase Transitions and Critical Phenomena

The $G(n, p)$ ensemble exhibits sharp phase transitions:

• Emergence of the giant component at $p = 1/n$.
• Connectivity at $p = (\log n)/n$.

A refined analysis is the Karp–Sipser process: recursive removal of leaves (degree 1 vertices and their neighbors) leaves a Karp–Sipser core. At the critical value $\lambda = e$, $G(n, \lambda/n)$ has a core of size $O(n^{3/5})$, rigorously described by a diffusion approximation and a random scaling constant determined by a Brownian hitting time (Budzinski et al., 2024). Below criticality, the core is finite; above, it is a positive fraction of $n$.

Table: Karp–Sipser Core Size Regimes (Budzinski et al., 2024)

Regime ($\lambda$)	Core Size	Main Feature
$\lambda < e$	$O(1)$	Finite Poisson sum
$\lambda = e$	$O(n^{3/5})$	Brownian scaling, critical window
$\lambda > e$	$C(\lambda)n$	Positive fraction, central limit theorem valid

4. Spectral Structure and Universality

For $p = c/n$, the limiting empirical spectral distribution (ESD) is the spectral measure of a Poisson($c$)–Galton–Watson tree (Benjamini–Schramm limit) (Jung et al., 2017). As $c \to \infty$, the rescaled spectrum converges to the Wigner semicircle law,

$$\rho_{\mathrm{sc}}(x) = \frac{1}{2\pi}\sqrt{4 - x^2} \cdot 1_{|x| \leq 2}$$

and eigenvectors become delocalized: a $1-o(1)$ fraction have $\ell^\infty$ norm $o(1)$ (Jung et al., 2017).

In denser regimes ($pN \gg N^{2/3}$), the local semicircle law holds: the spectrum matches the semicircle law on spectral windows down to $N^{-1}$ (Erdős et al., 2011). Bulk universality (sine kernel for eigenvalue spacings) and edge universality (Tracy–Widom law for extremal eigenvalues) hold for appropriately rescaled adjacency matrices (Erdos et al., 2011).

5. Large Deviations and Rare Events

The large deviation principle (LDP) for $G(n,p)$ in the dense regime describes the exponential decay of probabilities of rare events (e.g., large subgraph counts) in the topology of graphons: $$I_p(f) = \iint_{[0,1]^2} \left[ f(x,y)\log\frac{f(x,y)}{p} + (1-f(x,y))\log\frac{1-f(x,y)}{1-p} \right]\,dx\,dy$$ (Chatterjee et al., 2010). This variational principle underlies upper-tail probabilities for subgraph counts, leading to “replica symmetry breaking” where the typical graph conditioned on a large subgraph count is either homogeneous or features a planted clique, depending on the subgraph density (Chatterjee et al., 2010, Cook et al., 2018).

In the sparse regime ($n^{-1/(2\Delta)} \ll p \ll 1$), sharp large-deviation estimates for subgraph counts are established via quantitative Szemerédi-type regularity and counting lemmas, with variational problems encoding the rate (Cook et al., 2018).

6. Algorithmic and Statistical Aspects

In group-testing analogues, learning the edge set of $G(n,q)$ by non-adaptive set queries can be achieved with $O(q n^2 \log n)$ tests and decoding time nearly linear in the expected number of edges. Binary splitting and partitioning strategies are shown to achieve the fundamental information-theoretic limits with efficient decoding (Ta et al., 21 Nov 2025).

In detection and matching problems for correlated $G(n,q)$ graphs, the “low-degree” method shows that any polynomial-based test of degree $O(\rho^{-1})$ (where $\rho$ is edge correlation) cannot reliably distinguish between correlated and independent pairs for small $\rho$. In the sparse regime, the detection threshold is given by a value $\sqrt{\alpha}$, where $\alpha \approx 0.338$ is Otter’s constant, and no polynomial-time method based on subgraph counts can push the threshold below this (Ding et al., 2023).

7. Dynamics, Extensions, and Generalizations

Two classes of dynamic Erdős–Rényi models have been analyzed:

• Regime-driven: Edge appearance/removal rates depend on an external Markov process, leading to master equations and Ornstein–Uhlenbeck diffusion approximations for macroscopic statistics.
• Periodically-resampled: Edge update rates are periodically sampled; stationarity admits explicit moment formulas and functional CLTs (Mandjes et al., 2017).

Limit laws are obtained for the evolving number of edges, with large-deviation principles governing excursions, and functional CLTs for fluctuations. The tools extend to large-scale time-dependent random environments and provide a rigorous stochastic-process perspective.

---

Erdős–Rényi graphs, in both the finite and infinite setting, are the universal template for random structures in discrete mathematics. They exemplify concentration, phase transitions, universality in statistical physics, algorithmic complexity, and sharp large-deviation principles. A rich body of advanced research continues to refine their critical behavior, inferential limits, spectral properties, and dynamical analogues in ever more general settings (Cameron, 2013, Alarfaj et al., 2023, Budzinski et al., 2024, Jung et al., 2017, Erdős et al., 2011, Erdos et al., 2011, Dhara et al., 2016, Cook et al., 2018, Ta et al., 21 Nov 2025, Ding et al., 2023, Chatterjee et al., 2010, Mandjes et al., 2017).